# Phase 3: Staggered DiD — Interpretation Notes

## Executive Summary

The staggered DiD produces a nuanced picture: capital account opening has a **delayed positive
effect on current accounts** for transition economies, but **does not amplify the demographic
channel**. The demographics-CA relationship is **stronger pre-opening than post-opening**,
confirming the Phase 4h finding with modern methods.

---

## Key Results

### 1. TWFE Negative Weight Diagnostic (A4)
- **51.1% negative weights** in standard TWFE — severely contaminated
- Standard TWFE coefficient on post_opening is unreliable
- This validates using modern robust estimators

### 2. Triple-Difference: Does Opening Amplify Demographics? (A2)

**All openers (N=2,577):**
- Z₁ pre-opening: 56.5 (p=0.006***)
- Z₁ × post: -38.2 (p=0.106)
- Z₁ post-opening (net): 18.3
- **Demographics are WEAKER post-opening**, though the interaction is marginally insignificant

**CCA only (N=290):**
- Z₁ × post: -154.0 (p=0.083*)
- **Marginally significant in the WRONG direction** — opening REDUCES the demographic effect

**Transition economies (N=656):**
- Z₁ × post: -123.9 (p=0.049**)
- **Significant at 5%** — opening reduces the demographic coefficient by 2.4×

**Interpretation:** Capital account opening does NOT activate the demographic channel.
If anything, it WEAKENS it. This is the opposite of the lifecycle/openness-gating hypothesis.

### 3. TWFE Event Study (A3)

**All openers:**
- Pre-treatment coefficients: noisy but no clear trend (joint p=0.127, passes)
- Post-treatment: no significant effect in any period
- Pattern: CA does not systematically change after opening

**CCA only:**
- Very wide confidence intervals (SE ~ 3.5pp for each coefficient)
- No significant pre-trends (joint p=0.483, passes easily)
- No significant post-treatment effects
- Pattern: CCA opening has no detectable average effect on CA

### 4. BJS Imputation Estimator (Part B)

**All openers (N=2,965 treated obs):**
- Overall ATT: +0.10pp (p=0.69) — **no effect of opening on CA**
- No significant effect in any individual event-time period

**CCA only (N=162 treated obs):**
- Overall ATT: +3.59pp (p=0.10) — **marginally positive** (CA improves after opening)
- Individual periods: large point estimates (6-10pp at e=4 through e=11) but very wide SEs
- Interpretation: CCA countries tend to improve CA after opening, but imprecisely estimated

**Transition economies (N=433 treated obs):**
- Overall ATT: +5.35pp (p<0.001***) — **highly significant positive effect**
- Pattern: effect emerges at e=4-5 (4-5 years post-opening) and grows through e=13
- Strongest effects at e=9 through e=13 (8-9pp)
- **This is the DELAYED ACTIVATION pattern** — transition economies improve their CA
  substantially 5-15 years after opening, but not immediately

### 5. Pre vs Post Split Sample (C3)

**Pre-opening openers (N=236):**
- Z₁ = 70.8 (p=0.079*) — demographics matter before opening

**Post-opening openers (N=1,753):**
- Z₁ = 7.9 (p=0.689) — demographics become insignificant after opening

**Coefficient ratio: 70.8 / 7.9 = 9:1 pre:post**
This confirms the Phase 4h finding: demographics predict CA BEFORE opening, not after.

### 6. Delayed Activation (C4, C5)

**Level effects (C4):**
- None of the period dummies are significant (all p > 0.16)
- Weak hint of growth at 10-14 years (+0.95pp, p=0.16) — not reliable

**Z interaction effects (C5):**
- Z₁ × post_15plus: -59.9 (p=0.069*) — if anything, the demographic effect
  WEAKENS over time after opening, not strengthens

### 7. Cohort-Specific ATTs (Part D)

**Highly heterogeneous across cohorts:**
- 1996 cohort (incl MNG): -3.97pp (p=0.001***)
- 2002 cohort (incl AZE): +6.95pp (p<0.001***) — driven by oil boom
- 2008 cohort (incl TJK): +7.14pp (p<0.001***)
- 2012 cohort (GEO): +6.35pp (p<0.001***)
- 2016 cohort (KGZ): -7.70pp (p=0.199) — negative but noisy

**CCA cohorts mean ATT: +1.75pp** (positive but driven by commodity/recovery)
**Non-CCA cohorts mean ATT: -1.12pp**

**Aggregated ATT: -0.57pp** (insignificant) — no average effect of opening

---

## Synthesis: What Does This Mean for Causality?

### Evidence AGAINST the causal/openness-gating interpretation:
1. Triple-diff shows opening WEAKENS the demographic effect (transition: p=0.049)
2. Split sample: 9:1 ratio pre:post in demographic coefficients
3. Z × delayed activation interactions negative (wrong sign)
4. Average treatment effect of opening on CA is zero

### Evidence that SOMETHING is happening:
1. Transition economies show +5.35pp ATT on CA (p<0.001) — delayed, at 5+ years
2. CCA cohort-specific ATTs are heterogeneous but mostly positive
3. The delayed improvement pattern is real — countries DO improve after opening

### Reconciliation:
The most likely interpretation is that **capital account opening triggers structural reforms
and institutional changes** (banking development, FDI inflows, fiscal discipline from
international markets) that improve current accounts. But this improvement is NOT operating
through the lifecycle/demographic savings channel. Instead, the demographic-CA correlation
PREDATED the opening (reflecting post-Soviet structural factors) and was DISRUPTED by the
institutional changes that opening brought.

In other words:
- Before opening: demographics correlated with CA through structural/confounding channels
- After opening: those structural channels are disrupted by reform and integration
- The net effect: CA improves, but not because of demographics

### Implications for the paper:
1. **The causal story is not straightforward** — demographics predict CA but opening
   doesn't amplify this prediction
2. **The delayed positive ATT for transition economies is the strongest finding** — worth
   reporting as evidence that opening improves CA outcomes, even if not through demographics
3. **The pre>post coefficient pattern is consistent across specifications** — this is robust
4. **51% negative weights in TWFE** is a important methodological finding — validates using
   modern estimators
5. **Cohort heterogeneity** suggests country-specific analysis (Phase 4 SCM) may be more
   informative than pooled estimates
